SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillslongform-animated-talking-head

longform-animated-talking-head

General & other

View repository

Updated 10-15 23:45

About this skill

Long-form animated talking head (landscape 1920x1080)

Built reverse-first on the JEV AI x HyperEdit intro (2026-09-21, Kevin: "ya really good. solid long form skill start"). Reference cut = clone-projects/jev-intro-longform-animated/render.py (historical, not bundled) on top of templates/longform_engine.py. Same visual language as the short-form flagship (Oswald + Luckiest Guy, graphite wash, neon/orange/red + brand accents, real marks, GIFs, light fx, rare punches) but the canvas, layouts and hook rules are different.

Locked rules for long-form (Kevin 2026-09-21)

  1. SPLIT is the default layout for talking-to-camera (Kevin, screenshot of the CapCut/Premiere frame: "when i am talking full screen i like this editing layout you did the most"). Animation slot 1140x950 left, his head in the ringed 680x950 card right. FULL (face full-bleed + side overlays) only for the hook, one dramatic claim, the CTA. HERO (full-width slot + head PIP) only when a grid / thumbnails / b-roll IS the content.
  2. No mystery cold open. Short-form gets the dark wash + "?"; long-form does NOT. Frame 0 is a bright billboard on the ACTUAL topic (brand badges + wordmark + "the biggest update yet" chip beside the untouched face). Never darken the hook. No "?" badges later either.
  3. No bottom strip. Progress bar / label / timecode / chapter name were "truly pointless". Accent line at y=1038, captions rest on it, nothing under it.
  4. Head never moves. No punch_zoom / shake on the head in any layout. Slot layer only.
  5. Captions are phrase groups (<=4 words, Luckiest Guy 60px, white, 7px black stroke, active word in KEY_COL, unspoken words dim STEEL), baseline resting on the accent line.
  6. Consistent palette + type across a series. Pick brand accents once (PINK = Jev, VIOLET = HyperEdit; CYAN / NEON / RED per section) and reuse them every episode.

Layouts (engine constants)

SPLIT  (default) slot 1140x950 at (40,46) LEFT + head card 680x950 at x=1200 RIGHT (1:1 HEAD_CROP from
       meta, r=44, accent ring). Slot-local px: title cy≈70, left cx≈330, right cx≈850, chips y≈700, lane y≈800.
FULL   head full-bleed 1920x1080 + transparent overlay 1920x1080, bottom gradient under the captions.
       Author overlays on the RIGHT (x 1300..1900) or a top wordmark (cy≈84); the face stays clean.
HERO   slot 1840x950 full width + head PIP 520x292 at (1330,690) with ring. Keep hero pieces out of
       the bottom-right 520x292; lanes may run behind the PIP.

Switch layouts on spoken anchors (every ~4–8 s keeps it action packed). Layout change = 0.24 s zoom-dissolve (outgoing view rendered live at the same t, incoming eases 1.035→1.0 scale); same-layout section cuts whip. No shake and no hyperframe flash on any section cut (Kevin 2026-09-30: "I don't like the video shake transition from a different view"). Shake only fires on extra_hits impact beats.

Pipeline

  1. Workdir + prep (clip must exist on disk; landscape 1920x1080 expected, others cover-crop):
bash "${CLAUDE_SKILL_DIR}/scripts/setup_workdir.sh" <slug>        # projects/<slug>-longform/
cd projects/<slug>-longform
python3 "${CLAUDE_SKILL_DIR}/scripts/prep_longform_head.py" --clip "/abs/clip.mov" --whisper

Writes source/talking_head.mp4 (h264 30fps), audio/narration.{wav,m4a}, audio/words.json, source/meta.json (DUR, HEAD_CROP = 680x950 face-anchored crop for the SPLIT card) and renders/head_sheet.jpg. View the sheet: crown and chin inside the card, face ~42% down; override --crop-x/--crop-top if he drifts. Fix Whisper mishears in words.json first ("Jev"/"Jeff", product names) — captions, anchors and number beats all read it.

  1. Read the whole transcript with timestamps and plan sections at spoken anchors (tabs("word", after) / tabs_any([...], after)). SPLIT by default; one FULL for the hook, HERO for the grid moments. "Action packed and fast paced": 6–10 live pieces in the slot, a beat on every named product / number / punchline, a GIF on every joke or flex.

  2. Source assets (before choreography; scripts live in the short skill, call by path):

    • logos: split-animated-talking-head/scripts/fetch_brand_asset.py → assets/logos/. Proven marks (jev, typesafe, hyperedit, obsidian, github, capcut, premiere, youtube, ffmpeg, python, openai, gemini, deepseek) sit in clone-projects/jev-intro-longform-animated/assets/logos (historical, not bundled) — copy them. Claude family = bundled invader/sphinx. Premiere = Wikimedia SVG → qlmanage -t -s 512. CapCut = s2 favicon (tool_badge skips its white punch).
    • GIFs: fetch_giphy.py --query ... --list, pick by title, --id <id> --slug <slug>; ≤4 s; eyeball frame 1. One per beat.
    • thumbnails Kevin drops in → assets/thumbs/thumb_<i>.png → thumb_card(i, w) (crops the YouTube card to the 16:9 image, adds a play badge). Screenshots he pastes live in /var/folders/.../TemporaryItems/NSIRD_screencaptureui_*/ — copy with find -exec cp (a plain cp on those paths fails in the sandbox).
    • figures: fetch_figure.py whenever a company / niche with famous faces is spoken. Build a contact sheet of everything and view it before authoring.

  3. Author render.py from templates/render_template.py: SECT, T_*, KEYWORDS, KEY_COL, bespoke cards, scene(), then run(SECT, scene, punch_words, extra_hits, ...). Engine cards: chip / strike_chip / big_text_card / stamp_card / wordmark / ghost_img / ticker / draw_code_rain / sparkle_badge / round_logo / tool_badge + x_mark / terminal_card (typed lines) / timeline_card (fake NLE, playhead + agent cursor) / workflow_card (nodes light up) / thumb_card / demo_card / whisper_line; plus media_marks (gif_card, logo_badge, topic_marks, figure_*, counter/odometer), hyper_edits (entrance(k), lanes / guarded_conveyor, punch_beats, whip, hyperframe), light_fx (light_hit, light_rays, bloom_orb, speed_streaks) and claude_marks (invader / sphinx). Beat t0 = spoken word (max 0.08 s lead via r(T)); cold-open atmosphere with place(..., -0.35, dur=0.01) so each section's first frame is packed; exit_at the previous big piece ≤0.06 s before the next hero; rotate entrance(k); punches = ONE hero word per section (>=8 s apart); extra_hits = the 3–5 biggest claims.

  4. Preview, then render:

PREVIEW="0.0,2.2,5.5,12.6,21.5,29.5,44.6,52.2" python3 render.py      # frames/out/f<n>.jpg
bash "${CLAUDE_SKILL_DIR}/scripts/render_parallel.sh" "Readable Name" 4
# → renders/final.mp4 (1920x1080, the clip's audio, crf 16) + ~/Downloads/Readable Name.mp4

Tile the preview frames (3 columns at 960x540) and LOOK for: pieces under the head card / PIP, empty halves (add a terminal, lane or badge), a FULL-layout title over the cap, stamps over GIFs. Kevin reviews from ~/Downloads; keep the project for re-renders.

Learnings

  • 2026-09-21 v1 → v2 (JEV intro): dark hook + "?" = wrong for long-form ("it just needs to hook the user into the actual topic"); bottom timeline strip = "truly pointless". Both removed; slot/head card grew to 950 tall. Approved: layout switching, phrase captions on the line, X-marks over CapCut/Premiere on the spoken word, GitHub fork grid on "hundreds", thumbnails on "three video updates", Jev slam + rays, fake timeline, workflow nodes, code tree, demo card + "INSANE" stamp. Then: SPLIT named the favourite layout for talking.
  • HEVC .mov from Kevin's recorder: transcode to h264 30fps first (cv2 reads sequentially, seeks only at a RANGE start). 1572 frames render in ~3 min across 4 chunks.
  • Long-form Whisper: small.en (base mishears more on a long clip).
  • 10-minute clips: ~60–90 sections; ~0.15 s/frame/process → 18k frames ≈ 12 min on 4 chunks.
  • 2026-09-25 VIKTOR intro (examples/viktor-intro-longform/render.py): Kevin's raw clip had logo SLATES baked in (black frame + product logo, no face) at 35.4–37.9 s and 75.5–78.0 s → the head card rendered black. Scan every source for dark runs first (frame[::8,::8].mean() < 40) and put those spans on "SOLO" (any non-SPLIT/FULL/HERO name = full-width slot, no head, no PIP); boundaries must land on the first real frame. Name-drop reveal + five-fields board went there. Also: wordmark() clips tall Oswald digits / its sub overlaps at big sizes → use a textbbox-sized big_num() + a chip; Oswald has no ✓ glyph (Menlo does, so terminals are fine); white-bg Wikimedia logos need a white-trim crop before logo_badge(punch=False, fill=WHITE).